More Specific Announcements in BGP Geoff Huston APNIC

What’s a “more specific”? A prefix advertisement that refines a “covering” advertisement 10.0.0.0/8 10.1.0.0/16

Why advertise a more specific? I: To redirect packets to a different network: “hole punching prefixes” 10.0.0.0/8 AS 65530 Different Origin AS 10.1.0.0/16 AS 138000

Example: Type I Network Path >* 72.249.184.0/21 4777 2497 3356 36024 >* 72.249.184.0/24 4777 2497 2914 40824 394094

Why advertise a more specific? II: To redirect incoming traffic to different network paths: “traffic engineering prefixes” 10.0.0.0/8 AS 65530 10.128.0.0/9 AS 65530 10.0.0.0/9 AS 65530 Same Origin AS, different AS Path AS 65536 AS 65537

Example: Type II Network Path *> 1.37.0.0/16 4608 1221 4637 4775 i *> 1.37.27.0/24 4608 1221 4637 4837 4775 i *> 1.37.237.0/24 4608 1221 4637 4837 4775 i

Why advertise a more specific? III: To prevent more specific prefix hijacking: “more specific overlays” 10.0.0.0/22 AS 65530 10.0.0.0/24 AS 65530 Same Origin AS, same AS Path 10.0.1.0/24 AS 65530 10.0.2.0/24 AS 65530 10.0.3.0/24 AS 65530

Example: Type III Network Path *> 1.0.4.0/22 4608 4826 38803 56203 i *> 1.0.4.0/24 4608 4826 38803 56203 i *> 1.0.5.0/24 4608 4826 38803 56203 i *> 1.0.6.0/24 4608 4826 38803 56203 i *> 1.0.7.0/24 4608 4826 38803 56203 i

How many eBGP route advertisements are more specifics? AS 131072 – 7 June 2017 Routes Advertised Address Span eBGP Routes: 671,659 2.84B /32s More Specifics: 357,372 (53%) 0.82B /32s (28%)

Has this changed over time? Total Advertisements More Specifics IPv4 2017 2007

Has this changed over time? IPv4 IPv6

Has this changed over time? 38% 50% IPv4 IPv6 Ratio of More Specifics : Total Advertisements (%)

Has this changed over time? Surprisingly not for IPv4! In the IPv4 network more specifics have been ~50% of the Internet’s announced route set for the past 10 years. In IPv6 the relative number of more specifics is climbing, and now stands at 40% of the total set of announced IPv6 prefixes

More Specific Types – Prefix Counts

More Specific Types – Prefix Counts IPv4 IPv6

More Specific Types – Relative Counts IPv4 IPv6

More Specific Types In both IPv4 and IPv6: Type I prefixes (”hole punching”) are declining over time (relatively) Type II prefixes (“traffic engineering”) have been relatively constant at some 30% of more specifics Type III prefixes (“overlays”) have risen (relatively) and are now the more prevalent form of advertised more specifics in both IPv4 and IPv6

What about Address Spans covered by more specifics?

What about Address Spans covered by more specifics? Despite IPv4 address exhaustion from 2011, the span of addresses that are announced by more specifics continues to grow by some 32M /32’s per year

What about Address Spans covered by more specifics? IPv6

What about Address Spans covered by more specifics? 26% 2.5% IPv4 IPv6 Ratio of More Specific Address Span : Total Advertised Span (%)

Address Span: Breakdown into Types IPv4 IPv6

% of Total Span: Breakdown into Types IPv4 IPv6

Overlays are the majority of More Specifics In both protocols the largest block of more specific announcements in terms of address span are “overlays” where the AS Path of the enclosing aggregate and the more specific are identical The initial IPv6 network had little in the way of overlays and had a high proportion of Type I (Hole Punching) more specifics. This has changed over time and the recent profile is similar to IPv4

BGP Updates Overlays do not change routing, but do they add to the routing load? Are more specifics “noisier” than aggregates? Are overlays more active in terms of BGP Updates than other more specific types?

Update count by Prefix Type

Update count by Prefix Type IPv4 IPv6

Update Count In IPv4 the update count for more specific prefixes is greater than the comparable count for root prefixes, while the opposite is the case in IPv6. But the relative count of more specifics is ten times lower in IPv6 Let’s “normalise” this by dividing the update count by the number of prefixes to get the average update count per prefix of each type

Relative Update count by Prefix Type

Relative Update count by Prefix Type IPv4 IPv6

Relative Updates On average, in IPv4 More Specifics are slightly noisier than Roots, while in IPv6 roots and more specifics are equally likely to be the subject of BGP updates Are different types of more specifics more or less stable in BGP terms?

Average Number of Updates Per More Specific Prefix Type

Average Number of Updates Per More Specific Prefix Type IPv4 IPv6

Average Number of Updates Per More Specific Prefix Type In IPv4 Type II Traffic Engineering Prefixes show a slightly higher level of BGP instability on average over Type I Hole Punching Prefixes, while Type III Overlay Prefixes show the lowest average update rate of more specifics In IPv6 this has only been apparent in the past three years, where Type II Traffic Engineering Prefixes are showing the greatest levels of BGP instability and Type I Hole Punching more specifics showing the lowest update rates

BGP Instability is heavily skewed IPv4 IPv6

BGP Instability is heavily skewed Instead of looking at update profiles averaged across all prefixes, lets now look only at those prefixes that showed instability (were updated) each day (“active” prefixes)

What Type of Active Prefixes are more Unstable?

What Type of Active Prefixes are more Unstable? IPv4 IPv6

What Type of Prefixes are more Unstable? Type II More Specific Prefixes (Traffic Engineering) are approximately twice as likely to be unstable than either root prefixes or other types of More Specifics in both IPv4 and IPv6 This matches a rough intuition about the nature of more specifics, where overlays and hole punching would be expected to be as stable as root announcements

Average Number of Updates per Active Prefix Type IPv4 IPv6

Average Number of Updates per Active Prefix Type In IPv4 Type II Traffic Engineering Prefixes have a greater average number of updates than other prefix types In IPv6 Root Prefixes tend to have a lower average number of updates than other prefix types Perhaps the significant message here is that IPv6 has a higher inherent level of routing instability – unstable prefixes in IPv6 have 100x more instability events per unstable prefix on average than unstable prefixes in IPv4

What if… All Overlay more specific prefixes were removed from the routing table?

Table Size Implications IPv4 IPv6

Update Count Implications IPv4 IPv6

What if… All Overlay more specific prefixes were removed from the routing table? Both IPv4 and IPv6 routing tables would drop in size by approximately 30%, as Overlay more specifics are now the predominate type of more specifics in the routing tables The rate of dynamic instability in BGP would not change by any significant amount, as overlay more specifics are relatively stable prefixes

Summary of Findings More specifics add to both the size and the update load of BGP

Summary of Findings More specifics add to both the size and the update load of BGP However BGP itself is both a reachability and a traffic engineering tool, and more specifics are often used to qualify reachability by traffic engineering. We have no other viable internet-wide traffic engineering tools, so this particular use of BGP really has no alternative

Summary of Findings More specifics add to both the size and the update load of BGP However BGP itself is both a reachability and a traffic engineering tool, and more specifics are often used to qualify reachability by traffic engineering. We have no other viable internet-wide traffic engineering tools, so this particular use of BGP really has no alternative Recent years have seen the decline of hole punching as more providers tend to treat their address blocks as integral units.

Summary of Findings More specifics add to both the size and the update load of BGP However BGP itself is both a reachability and a traffic engineering tool, and more specifics are often used to qualify reachability by traffic engineering. We have no other viable internet-wide traffic engineering tools, so this particular use of BGP really has no alternative Recent years have seen the decline of hole punching as more providers tend to treat their address blocks as integral units. Overlays are becoming more prevalent as a means of protecting a routing block from more specific hijack threats. While this has implications in terms of total table size it has no significant impact on BGP update rates.

Summary of Findings More specifics add to both the size and the update load of BGP However BGP itself is both a reachability and a traffic engineering tool, and more specifics are often used to qualify reachability by traffic engineering. We have no other viable internet-wide traffic engineering tools, so this particular use of BGP really has no alternative Recent years have seen the decline of hole punching as more providers tend to treat their address blocks as integral units. Overlays are becoming more prevalent as a mean of protecting a routing block from more specific hijack threats. While this has implications in terms of total table size it has no significant impact on BGP update rates. There is the question of total instability in IPv6 being far greater than IPv4, but this is not intrinsically an issue with more specifics, but a more general issue of BGP routing instability in IPv6

Thanks